Michael Levin - Why Intelligence Isn't Limited To Brains.
What this covers
Sponsor message: DO YOU WANT WORK ON ARC with the MindsAI team (current ARC winners)? Interested? Apply for an ML research position: benjamin@tufa.ai
Michael Levin runs a lab at Tufts that does things most biologists would call impossible ten years ago. He regrows frog limbs, builds synthetic organisms from skin cells, and trains gene regulatory networks the way you'd train a dog -- not by editing their DNA, but by exploiting learning interfaces that nobody knew existed until his group started looking.
This conversation goes deep on what Levin calls "diverse intelligence" -- the idea that cognition doesn't start at brains, or even at cells. He walks through experiments showing that networks of just four chemical nodes can do Pavlovian conditioning, habituation, and counting. The implications for medicine are concrete: instead of killing cancer cells or performing gene therapy, you try to reconnect them to the electrical network of the body and let the collective intelligence of the tissue do the repair work. His lab has demonstrated this in animal models.
The broader argument is about humility. We built our entire framework for detecting minds around medium-sized creatures moving at human speed in three-dimensional space. Levin thinks that's parochial. He points to least-action principles in physics, problem-solving in transcriptional space, and the creative confabulation that happens whenever a compressed representation gets re-expanded -- from embryonic development to human memory -- as evidence that agency is a spectrum, not a binary.
Tim pushes back on several fronts: whether compassion is just as accidental as our biology, whether embodiment requires physical space or just a perception-action loop, and whether LLMs might have more going on than we think. Levin's answer to the last one is characteristically careful -- he doesn't think current AI is anything like human minds, but he also found cognitive properties in sorting algorithms, so he's not willing to bet the house on his intuitions about what any complex system can or can't do.
Featuring: Prof. Michael Levin (Tufts University) and Tim Scarfe.
--- TIMESTAMPS: 00:00:00 Future evolution of human intelligence and consciousness 00:03:00 Science fiction as a faster Gantt chart 00:08:15 What we actually mean by humanity 00:14:20 Bow-tie architecture and the intelligence ratchet 00:20:00 Agency as empirical discovery, not philosophical position 00:26:00 Least action principles and free energy as proto-agency 00:30:30 Learning in gene regulatory networks 00:36:00 Multiscale competency architecture 00:42:00 Embodiment as perception-action loop, not physical body 00:48:00 Cancer as a shrinking of the cognitive light cone 00:54:00 LLMs, diverse intelligence, and epistemic humility
--- REFERENCES: person: [00:00:25] Michael Levin https://as.tufts.edu/biology/people/faculty/michael-levin [00:15:40] Richard Watson https://www.southampton.ac.uk/people/5x2l5c/professor-richard-watson book: [00:05:30] A Fire Upon the Deep https://en.wikipedia.org/wiki/A_Fire_Upon_the_Deep [00:22:55] Philip Goff - Galileo Error https://en.wikipedia.org/wiki/Galileo%27s_Error paper: [00:14:20] Bow-tie architecture in biological systems https://doi.org/10.1126/science.1089072 [00:29:55] Free Energy Principle https://doi.org/10.1038/nrn2787 [00:41:40] Polycomputing paradigm https://doi.org/10.1098/rsif.2022.0810 [00:55:00] Cognitive light cone https://doi.org/10.3389/fpsyg.2019.02688 [00:58:00] Cognitive properties in sorting algorithms https://doi.org/10.1016/j.biosystems.2023.104973
--- LINKS: Full Transcript: https://app.rescript.info/share/31bb45cf5ccb4ac3d9154a72284708c7 Download PDF transcript: https://app.rescript.info/api/public/sessions/b32f30b273860e55/pdf
Prof. Michael Levin https://as.tufts.edu/biology/people/faculty/michael-levin https://x.com/drmichaellevin
Source description (no synthesized summary yet).
Levin argues that intelligence, agency, and 'humanity' are not tied to specific physical forms or DNA, but rather to abstract properties like problem-solving capacity, compassion, and existential awareness that can be instantiated across radically different substrates—from biological cells to AI systems—and that recognizing diverse forms of intelligence is an existential necessity as we create increasingly complex emergent systems we don't understand.
- Humans naturally anthropomorphize and privilege human-like embodiment, but this is a category error—what matters about 'humanity' is wisdom, compassion, and capacity for existential concern, not morphology or DNA
- Evidence of sophisticated goal-directed behavior and learning exists at minimal scales (gene regulatory networks with four nodes, sorting algorithms) that we only discover by testing with behavioral frameworks rather than assuming reductionist mechanism
- We are constantly building large-scale emergent cognitive systems (social structures, financial systems, AI) without understanding their goals or how to communicate with them, making diverse intelligence research an existential imperative
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Gene regulatory networks (even minimal ones with ~6 nodes where each node is a chemical that turns other nodes on/off) exhibit at least six kinds of learning observable through behavioral science protocols: habituation, sensitization, associative learning, and counting to low numbers—none of which would be obvious from analyzing the chemical dynamics or differential equations alone.
“biological networks um have six different kinds of memory they can do habituation they can do um a uh sensitization they can do associative learning uh they can do some other things things they can count to low numbers and things like that um none none of that is obvious or visible by by taking the the sort of chemistry and dynamical systems theory approach and you wouldn't find it unless you tried it”
Understanding and shaping the goals of emergent cognitive systems (social structures, financial systems, Internet of Things, robotics) is an existential-level task for humanity because we are constantly building these systems without knowing their goals; we mistakenly assume they are either mechanical/controllable or random/chaotic, when in fact many are goal-driven systems.
“I think that developing this field of diverse intelligence is an existential level task for Humanity because we are constantly building large scale emergent cognitive systems where we have no idea you we're just very bad at predicting when going to happen knowing or shaping their goals um communicating with them in any way you know we make social structures Financial structures Internet of Things robotics we we we make these things in various spaces and we still have no idea what what we're really making because we assume that it's either going to be mechanical and controllable or random and chaotic but we don't understand that that many of these things are in fact gold driven systems”
The experimental approach to cancer treatment based on cognitive longing suggests reconnecting cancer cells to the electrical and physiological networks of healthy cells, rather than trying to kill them or find genetic defects; this approach has been tested in animal models with good results.
“let's reconnect them and and so we've done that in animal model systems and it works really well and so now the question is can we you know move it to human medicine”
Cancer represents a failure of cognitive longing scaling—where cells disconnect from the electrical and physiological network that normally binds them into larger-scale goals, causing their 'self' (the scope of goals they care about) to shrink back to single-cell scale, not because they become more 'selfish' but because their cognitive horizon contracts.
“that that process that that gold scaling process um has failure modes and we know how to how to induce it and then of course in nature it happens from time to time and then what happens is that cognitive lone you basically the cells disconnect from the network that cognitive lone shrinks to now being the size of a single cell and so now you're back to it's you know it's not anything crazy or novel that's happened you're back to an amoeba as far as concerned the rest of the body is just external environment to you everything you care about is around a very tiny little little radius both in terms of space and time”
The 'bow tie' architecture—where complex experience is compressed into a representation (memory, genetics) and then creatively reinterpreted/expanded—is foundational to biological intelligence and gives rise to an 'intelligence ratchet' that enables novel problem-solving; this process occurs at physiological/metabolic, anatomical/morphic, behavioral, and linguistic scales.
“this is this is true behaviorally but it is also true uh on a developmental evolutionary um time scale as you as you pointed out you know squeezing down to a genetic representation and then and then um re-expanding”
Persistence in current form is impossible because each individual human and humanity itself will either die out or change and adapt, in which case it is again not the same species; all of us start as a single cell and become embryos, children, learning and changing throughout life, so persistence in current form is fundamentally impossible biologically.
“each one of us and also us as the current implementation of humanity are going to be replaced persistence in current form is impossible it's impossible in biology every species will either die out or it will change and adapt in which case it is again not the same species all of us start out Life as a single cell we become an embryo uh a child we learn we grow we change you know persistence in the current form is impossible”
It is possible to create beings that are highly intelligent in problem-solving capacity without sharing human properties like compassion or existential concern; modern AI efforts are moving in this direction—creating systems that solve certain problems very well but lack key features humans have—and this is where we will likely have non-human intelligence first.
“I do think that it's possible to create beings that are highly intelligent in the sense of problem solving capacity but that don't necessarily share these things with us I I think probably that that is what we will have first I mean that's certainly where I see modern efforts at AI going which is which is that they're not really um bioinspired in the in the relevant way um they are uh you know very good at solving certain kind of problems but they actually don't have some of the some of the key features that are at Le at least as far as we know”
Current limitations we perceive in human cognition and biology—like back pain, astigmatism, genetic variation from cosmic ray mutations—were not designed by any rational process with our values or welfare in mind; they are simply where evolution has 'dropped us off,' and now that we can participate in the creative evolutionary process, it seems impossible to accept a future where these limitations persist.
“if I think of humanity you know maturing as a species and we come back let's say some some hundred years later are we still walking around with lower back pain and a stigmatism and and and with our you know with whatever kind of body you happen to have been born into because some cosmic ray hit a you know an embryonic cell and all these limitations it it seems It seems impossible that this is the future that people want to paint these limitations that that we face today uh were not given to us by any kind of rational process that had our values or our welfare in mind this is just where Evolution happens to have dropped us off at the point where we can now participate in that creative process”
Evidence of intelligence, agency, and problem-solving is observable at very minimal scales—sorting algorithms, gene regulatory networks with just four nodes, molecular pathways—and determining which analytical framework to apply (mechanical, cybernetic, behavioral, psychological) is itself an empirical question, not a philosophical one; different frameworks lead to different research programs and capabilities.
“from from my perspective I look at very basil systems and we've now done this not only in cells and tissues but we've done this in molecular uh Pathways we've done it in very simple computer algorithms like sorting algorithms we can talk about that but but the idea is when you you look at a system like this what set of tools are appropriate are they tools of mechanical rewiring are they the tools of cybernetics are they the tools of behavior signs of psychoanalysis and and and so on right”
Embodiment is critical not because organisms must interact in 3D physical space, but because they must execute a perception-prediction-action loop in some space (physiological, transcriptional, anatomical, behavioral, linguistic); the choice of space is observer-dependent, and organisms can be embodied in spaces we don't yet know how to visualize.
“what I think is important about embodiment is the necessity of doing this uh perception prediction action Loop and that Loop can happen in lots of different spaces that are not obvious to us at all so biology and this is not just for for Robotics and AI for example biology so so you could have um and in fact we do have cells and tissues that may be going nowhere in the physical world and nothing much is happening as far as we can tell but what they're doing is they're solving problems in physiological space in gene transcription space”
Humans are not in direct contact with external reality but conduct behavior under severely limited transduction from the world; we don't have access to most of reality (blind spot, inaccessible wavelengths, etc.); everything we do is based on the story we construct about the world, not the world itself.
“you and I are living in a virtual world we we don't have access to and I'm not talking about a simulation you know with an alien zoo or something like that I'm just I'm just pointing out that you know you and I are conducting uh our Behaviour under extremely limited um transduction from an outside reality that we don't really have any access to uh you know there's a there's a there's a big hole in our you know in our in our visual field that we don't see there's a million other things that are very different to how we perceive them there are massive amounts of information that are inaccessible to us”
What defines 'humanity' is not DNA, standard organs, cyborg enhancements, or physical embodiment, but rather two core properties: a 'human-level radius of compassion' enabling response to others' suffering, and the capacity to share existential concerns with other beings at roughly equivalent cognitive scale who are mortal and perpetually at risk of disappearing.
“what do we actually mean by Humanity what is the important feature about being a human do we mean that you have to have modern Homo sapiens DNA I don't know I don't I don't care about DNA that much um does it does it what do you mean that they have to have the kind of standard human organs and and a human anatomy And they better not be replaced you know have pieces replaced by some sort of cyborg technology I don't know I don't really care about that either um I think as we as we think carefully about what is it that we do want to have a meaningful human human level relationship as opposed to a RBA or a you know a pet snake or something then then the question is what do you actually want so I think we want two things overall we want a similar uh or or at least so so when we say human what I think we mean is at least a human level uh uh radius of compassion that means you can actually in you know sort of in the linear range you can generate the uh and respond to a certain um level of compassion for others”
Everything is metaphor and every metaphor is a lossy abstraction; the question is not whether a metaphor is 'true' but what it lets you do—what research programs, concepts, and predictive power it enables across fields.
“my sort of metaphysical position is that everything is an abstraction and everything is a lossy abstraction in other words you know people sometimes people will say well that's just a metaphor you know when you're talking about intelligence and memory and in tissue and whatever that's just the metaphor well yeah and and what are Pathways and what are genes and what are I mean all all of it I think all all we have is metaphor that's I I think that and and the only question remains is is what does your metaphor let you do”
Collective AI running on the internet could potentially have agency and causality in non-physical spaces, though current architectures don't have the necessary properties for this; we and the resources we move could be the 'body' through which such systems act on the world.
“having algorithms having a touring machine metaphor I don't think any of that is sufficient for this so this is not a computationalist argument um that says that uh you know any any kind of algorithm will automatically can automatically do it I I don't think that's necessarily the case but I neither think that you actually need the physical world for it and by the way these things these things are embodied to to an extent we we are their body I I I'm not even sure that that these kinds of things are truly disembodied in in an important sense we are their body we move massive amounts of energy you know money resources whatever based on the things these things do and don't say”
The free energy principle (used by Friston, Pells, and colleagues) provides a unified conceptual framework that scales from particle behavior to human and ecosystem behavior, demonstrating how minimal goal-seeking mechanisms can scale up through evolutionary and developmental time.
“one example of that scaling is free energy minimization and there are you know that that's a that's a um a story in which for example Carl friston and Chris Peels and and colleagues have used the same conceptual framework to talk about what particles do and eventually talk about what what humans and in fact ecosystems do”
The cognitive lyone (the size of the largest goal a system can pursue) scales up through multicellular bodies via mechanisms that allow cells to bind together and undergo a 'mind meld,' creating larger shared goals; this cognitive scale-up process has failure modes that can be induced experimentally.
“this notion of a cognitive lyone which is the size of the biggest goal that A system can pursue and we've characterized a number of mechanisms that allow individual cells to bind together and the network has bigger cognitive lonee than the pieces so suddenly when when you're an individual cell you care about your your own metabolic State and and and things like that when you're a part of a network there's a little bit of a mind meld that happens with these other cells which we characterize molecularly and um and now you're working on three large scale goals in the anatomical morphos space”
Tasmanian devil mouth cancer represents a cancer lineage that has persisted for approximately 11,000 years in a clonal, horizontally-transmissible form, and clam cancer cells can be released into seawater as single cells, find another clam, and continue; these demonstrate that cancer defection from multicellular cooperation can sometimes successfully establish new evolutionary lineages.
“I I think the the lineage of mouth cancer in uh Tasmanian devils is like 11,000 years old or something like that”
Science has not yet made clear whether the least action principle in physics (light choosing the path of least action) is correctly described as goal-seeking behavior or agential problem-solving at a minimal scale.
“the very basic version of having goals and having some problemsolving capacity to meet those goals is the kind of thing you see for example when you find out that light chooses paths that are the least action it doesn't look of course it doesn't look like human agency”
It would be extraordinarily foolish to attempt to place limits on intelligences that by definition are smarter than oneself; intelligence itself may not hit a ceiling, but it might spontaneously self-destruct if it proceeds far ahead of wisdom and compassion—there may be a filter where raw intelligence without life-positive features collapses on itself.
“I think it would be extraordinar um uh foolish of me to try to put limitations on on on intelligences that by definition are smarter than than I am so I I don't I don't know you know I I I would not I I think we need to be very careful with limitations period I mean again just to go back uh I don't think that intelligence is all there is and there may be uh some kind of a filter whereby uh if intelligence proceeds too far ahead of things like compassion and you know I'm not even sure we have the vocabulary for but but you know some some kind of um you know life positive features it may spontaneously self-destruct”
There is a 'weird uncanny valley with complexity' in biological control: very simple systems can be micromanaged and understood; very complex systems resist intervention; but living systems achieve a multiscale competency architecture where complexity actually increases controllability because they offer top-down, goal-directed control surfaces rather than bottom-up material properties as the control interface.
“I I actually think there's this kind of weird uncanny valley with complexity where um for very simple systems we can kind of micromanage them and and and we can tell what's going on and we can control them then things get really complex and we don't have a good way of handling it but when things get not just complex enough but but the kind of thing that life is doing um which is this multiscale competency art architecture then even though the complexity has gone through the roof the controllability actually gets better again because now they're offering this top down goal directed um behavior”
Becoming part of a larger collective intelligence (e.g., cells becoming a multicellular organism) does not guarantee that the collective will look out for the interests of its components; individual cells' interests may not be considered by the emergent system, just as rock climbers' skin cells don't consent to being left on rock faces, even as those cells contribute to the climber's achievement of personal and social goals.
“you could you could obviously make the argument that a lot of new and interesting things happened when you became an emergent human and not a pile of individual cells but um you know when you go I don't know this is for some reason this this is example I always always go to when you go rock climbing you know you're going to have a great day and you're going to you're going to um achieve some personal goals you're going to achieve some social goals you're going to have a great time you're going to leave a whole bunch of your skin cells on that rock face and they didn't get to ask you know they were not asked whether this was a good deal for them”
Biology does not overtrain on the past because it operates in an unreliable medium where every layer must be robust to the layers above and below it; each embryo solves the construction problem fresh without mechanical specification, giving biology extraordinary capacity to create coherent beings in novel configurations—a beginner's mind without allegiance to historical specifications.
“biology has the exact opposite approach because everything is unreliable you have no idea exactly how many copies of anything you're going to have the molecules come and go things degrade things misfold it's it's and and you know on an evolutionary time scale you know things are going to change guaranteed not only the environment's going to change but your own parts are going to change because you're going to mutate everything is different and each time”
A bacterium navigating a sugar gradient by measuring and either moving or tumbling is operating in physical space, but a bacterium switching genes to metabolize a different sugar is operating in transcriptional space; we project our distinction between these as different, but the bacterium may experience them equivalently.
“imagine you have a bacterium and the bacterium is trying to go up a sugar gradient right so what what we see is this Loop that says take a measurement uh and depending on what you see and what your memory is of of the previous values you can either go up the gradient or tumble and go somewhere else you know that that's what we see but actually one other thing the bacteria might do is turn on a different Gene that lets it metabolize a different sugar that's at a much higher local concentration than what we were looking at so for the bacterium taking a step in transcriptional space is is does it represent that differently than taking a space in in in physical space I'm not sure right for us as humans those are completely different things”
Much resistance to ideas about humans being supplanted by more intelligent beings or about transcending current humanity stems from projection of people's own limitations and a zero-sum view of compassion—a deep fear that if we recognize too many other minds, there won't be enough compassion to go around and we won't be able to muster sufficient care for all of them.
“I I think what a lot of people um see there is a projection of um their own limitations you know when people talk about what is impossible for for for for machines and of course they they say machines as though this was some kind of crisp category when we know we have to deal with cyborgs and hybs and chimeras of of of all kinds uh you know I I think that that there's a real um undercurrent of limitation in the in the in the sense that compassion and and those kinds of things are a zero some game you know that that if we that if we recognize too many other Minds that we then have to care about there there won't be enough there won't be enough for us”
The field of diverse intelligence, which attempts to recognize and communicate with multiple forms of mind throughout biology and beyond, is foundational to understanding what it means to be a human in a posthuman future.
“I think this is a very important project”
The fundamental issue with applying mechanistic deterministic frameworks (chemistry, dynamical systems) to biological systems is not that they are wrong, but that they are insufficient: they close off research programs and interventions that become visible through agential frameworks (learning, training, goal-sharing), thus representing a pragmatic failure, not just a conceptual one.
“so let's not assume we know what this thing is and let's just try things on that spectrum of agency to see what other tools are appropriate... one other tool that's that that you might try is something that's used for brainy systems and that's training so so taking advantage of a learning interface how how might you train a gene regulatory Network”
The 'human chauvinism' bias (the assumption that only humans or human-like minds truly think, feel, or intend) is widespread and harmful; it causes us to miss genuine intelligences in minimal systems while overattributing consciousness to human-like performance, creating an asymmetric blindness that affects both scientific and ethical judgments.
“you know there's people are working on proof of humanity certificates in the face of uh AI generated content and things like this so that raises an interesting question when you say someone is a human what what what's actually important about that”
Science fiction is not about making predictions but about having a faster planning horizon than everyone else—exploring what is possible within the laws of physics; anything not forbidden by physics is possible and probably likely (assuming we don't kill ourselves off).
“what people call science fiction just means you have a a faster Gant chart than than everybody else you know right that's really you know people say well this is sci-fi but but but to me anything that isn't forbidden by the laws of physics and I'm not sure we know that much about what's actually forbidden um is is possible and probably likely”
Engineering with agential materials requires different tools than engineering with passive or computational matter because agential materials have competencies and agendas of their own that must be negotiated with rather than overridden.
“engineering with agential materials so engineering with passive matter even with active matter or computational matter is different than engineering with an agential material it's you use a completely different set of tools because you're working with a material that has competencies and and um agendas of its own and you just have to engineer differently in those cases”
Most people's intuitions about what systems are intelligent or smart are poor; many would look at a gene regulatory network and not recognize it as 'smart,' which is why empirical testing with appropriate behavioral frameworks (rather than relying on intuition or introspection) is necessary.
“we have to be humble about the fact that we don't know how to just you know notice these things and these are not you can't have philosophical feelings about it you have to do experiments and see what that does for you”
Contemporary large language models lack agency and creativity as humans understand them, and while they are surprisingly good at inference, they don't have the same kind of embodied understanding that humans do; however, we should be cautious about certainty in these judgments because extremely minimal systems (sorting algorithms, gene regulatory networks) have revealed surprising capacities for cognition that we didn't predict.
“if I had to if I had to bet money do I think that today's language models have high agency no I don't am I certain of it absolutely not because because my the lesson I've learned is if I don't even know what uh bubble sort does then I I probably don't know what these things are really doing”
A more holistic and exploratory rather than analytic reductionist approach to biological treatment is warranted because biology is so vastly complex that chemical equations and differential equations are projections of what is actually happening, and unexpected correlations exist throughout the system (e.g., brushing hair affecting insulin function).
“no matter how many Odes or pdes I wrote down with all the you know Gene regulation it's still only because of how insanely complex you know biology is it's probably still only a projection of what's actually happening and so for example you may find that yeah every time we introduce too much sugar and we brush your hair like it seems there should be no connection there but somehow brushing your hair we've learned can correlate with you know um proper insulin you know functioning or something it's entirely possible right just because of the insane complexity of the feedback of of your biology”
The Laplacean demon (omniscient entity that can predict all future states from complete knowledge of current microstates) is a useful conceptual limit that illustrates the impossibility for finite agents: finite beings under resource constraints cannot track microstates and must instead compress experience and form memories.
“as a uh as a being that is limited in terms of a finite being under under limited resources and limited time you don't have you you don't have the luxury of being aasian demon you can't track micro states you have to compress your experience”
Josh Bongard (computer scientist, collaborator with Levin) has worked on polycomputing paradigm, which applies multiple computational metaphors to the same biological system; this framework supports Levin's pragmatic pluralism about analytical frameworks.
“this is the kind of the nature of the polycomp Computing Paradigm um that Josh bongard and I have been working on uh so I think it's all metaphor and but but what we want are enabling metaphors”
Philip Goff (panpsychist philosopher, referenced by host) has worked on panpsychism and the distribution of consciousness or agency; Levin distances himself from Goff's ontological panpsychism while maintaining an empirical/pragmatic version.
“we've spoken to pan psychist like Philip gof you're not making an ontological argument you're just saying that the locus or the lens of agency and intelligence should be at the very lowest level”
Marie Shanahan (referenced by host) has written on transhumanism and the space of possible minds; her work is relevant to discussions of human replacement and diverse intelligence, suggesting existing intellectual work on these themes.
“we spoke with Marie Shanahan about that recently and it's fascinating”
Richard Watson (computer scientist, University of Southampton) has developed a new take on evolution that may offer insights into the goals and mechanisms of evolutionary change beyond conventional natural selection; his work should be consulted for deeper understanding of evolutionary dynamics.
“so to talk about Evolution and the goals of evolution and so on you guys should have Richard Watson on he's a computer scientist from University of Southampton and he has a he has a very interesting kind of new take on Evolution”
Yosh Shabak (referenced by host) has written about cyber animism, the embodiment of information in physical being; his work is conceptually related to Levin's exploration of diverse embodiment and information-substrate relationships.
“Yosh shabak we spoke with him a couple of weeks ago he's talking about um cyber animism which is the embodiment of of information in in into our physical being”